Dropbox Product Sense: Applying AI to Speed Up File Retrieval

As a product manager at Dropbox, how would you use AI to help people get to the files they need more quickly? Choose the users you would build for, explain why locating a file is slow for them today, propose AI-based solutions, decide what to ship first, and define how you would measure success. ```hint Diagnose before you reach for AI Locating a file fails in different ways: someone remembers what a document says but not what it is called, a teammate does not know a file exists, a photographer faces thousands of camera-named images. Pick the failure that matters most for your chosen users, then choose the AI capability that fixes it, instead of listing everything AI can do. ``` ### Clarifying Questions - Is the main goal retention and engagement of existing customers, growth of paid plans, or differentiation from competing cloud storage products? - Should I focus on individual accounts or on business teams, whose shared folders are created and named by many different people? - Are all file types in scope (documents, scans, images, audio, video), or mainly documents? - Should I build on the search Dropbox already offers, and what limits apply to processing file contents with AI models, such as user consent or team admin policies? ### What a Strong Answer Covers - A clear goal and a chosen user segment, with a reason that segment's file-finding problem is the most valuable to solve - Distinct ways that finding a file fails (forgotten name, content that is not text, many near-duplicate versions, unknown location), each tied to the AI capability that addresses it - A prioritized set of solutions weighed by user impact, effort and model cost, with a clear first release - Success metrics such as time to the right file and search success rate, plus guardrails for relevance and trust - Privacy, sharing-permission and accuracy risks of running AI over stored files, and how to handle them ### Follow-up Questions - How would you guarantee that AI search never surfaces a file, a snippet or a summary from something the user is not allowed to open? - Running models over every stored file is expensive. Which files would you process first, and would AI search be free or paid? - How would you design an experiment that shows the new search actually saves users time?